Multivariate time series classification (MTSC) remains a challenging task due to the limitations of existing methods in effectively capturing periodic patterns and extracting discriminative features. Although large language models (LLMs) have shown potential for time series analysis, their application is constrained by the modality gap between temporal data and language. In this work, we propose LLM4MTSC, a novel model that combines frequency domain analysis with text modality alignment to enhance the understanding of multivariate time series (MTS). Our method first transforms the time series data into frequency-domain representations using the Fast Fourier Transform (FFT), which decomposes the sequences into multi-scale subseries that explicitly capture periodic patterns. A subsequent temporal-text alignment module aligns the temporal and text modalities via a multi-head attention mechanism, integrating raw time series features with semantic embeddings to enrich the input context. Finally, LLM is employed to extract temporal features, which are combined with patch projection to obtain classification results. Experimental results demonstrate that our model outperforms state-of-the-art methods, achieving superior classification accuracy and effectively extracting the temporal features of multivariate time series.